seed-paper-profiler

name	seed-paper-profiler
description	Build a research-interest profile from a small set of seed math papers, abstracts, or user notes. Use when a coding agent should infer the user's target domain, core problems, methods, assumptions, proof patterns, technical keywords, negative preferences, and search directions before related-paper retrieval or paper-to-skill extraction.

Seed Paper Profiler

Purpose

Use this skill to turn a few seed papers into a research-interest profile. Do not summarize the papers as the final product. Extract a profile that can drive retrieval, triage, and later proof-pattern skill mining.

Inputs

Accept any mix of:

• Markdown papers such as paper.md.
• Abstracts, introductions, theorem statements, or user notes.
• Existing extraction reports or SkillCards from seed papers.
• Optional outputs/<research_project_id>/human_feedback_state.json if the user has already corrected or narrowed the research direction.

When a full paper.md is available, preserve source line references for claims used in the profile.

Workflow

1. Create or reuse outputs/<research_project_id>/.
2. Preserve seed inputs under seed_papers/ when writing artifacts.
3. Read each seed paper for problem setting, mathematical objects, assumptions, methods, proof patterns, evaluation criteria, and open questions.
4. Separate stable interests from one-paper accidents. Mark uncertain inferences as hypotheses.
5. Extract positive interests and negative preferences. Negative preferences include topics the user likely does not want to pursue.
6. Generate search directions covering problem, method, theorem, assumption, application, and adjacent-domain routes.
7. If human_feedback_state.json exists, apply its focus_updates, negative_preferences, and next_step_directives before finalizing the profile.
8. Write research_profile.json using the schema in references/schemas.md.
9. Write profile_review.md with a short checkpoint asking the user to confirm, remove, or emphasize directions.
10. When the user provides profile feedback, create or update human_feedback_state.json so later Skills can read the correction.

Language Policy

For this project, default to bilingual Chinese-English output for user-facing content.

• Write profile_review.md in Chinese and English. Use paired sections or paired bullets with 中文: and English:.
• When replying to the user after running the skill, summarize results in Chinese and English.
• Keep JSON keys and enum values in English for machine readability.
• For JSON natural-language fields, English is acceptable, but include checkpoint_questions_bilingual when asking the user to confirm the profile.
• Do not translate mathematical terms mechanically when a standard English term is clearer; pair it with a concise Chinese explanation when useful.

Required References

Read only as needed:

• references/schemas.md: required output fields.

Use the template:

• assets/research_profile.template.json

Output Files

outputs/<research_project_id>/
├── seed_papers/
├── human_feedback_state.json
├── research_profile.json
└── profile_review.md

Profiling Rules

• Prefer reusable research intent over paper-specific details.
• Track evidence for important claims when source lines are available.
• Keep user_interest_hypotheses explicit; do not present guesses as facts.
• Generate queries that can be used by the active agent's search or browsing tools.
• Include checkpoint_questions that help the user correct the direction before retrieval begins.
• Treat research_profile.json as the active agent's current interpretation and human_feedback_state.json as the user's correction layer.
• Suggested profile feedback should be concrete enough to affect retrieval, such as focus updates, down-ranked topics, or next-step directives.

Completion Check

Before finishing, confirm:

• research_profile.json is valid JSON.
• It includes domain, core problems, methods, assumptions, proof patterns, keywords, search queries, checkpoint questions, and bilingual checkpoint questions.
• profile_review.md is bilingual Chinese-English.
• Every high-confidence profile item has either source evidence or a clear note that it came from user-provided context.